Complaint Classifier
Sorts complaints into the Bank's taxonomy, with urgency and sentiment.
Macro-F1 (category)
0.872
Macro-F1 (Bahasa Malaysia)
0.861
Macro-F1 (Manglish)
0.834
Scam flag recall
0.95
1 more metric
Median latency
42 ms
About this model
Classifies customer complaints written in English, Bahasa Malaysia or Manglish into the Bank's complaint taxonomy (aligned to BNM complaint categories), with product, issue, sentiment and urgency. Flags potential scam or unauthorised-transaction cases for same-day handling and routes each complaint to the owning product team.
Intended use
Routing and prioritising complaints from branches, the contact centre, email, the app and social media; complaint trend reporting. A complaints officer confirms category on every case that is reported externally.
Training data lineage
Fine-tuned multilingual encoder (XLM-RoBERTa base) on 86,000 labelled complaints from complaints-corpus in BigQuery, with a held-out set double-labelled by the Customer Experience team. Trained and registered in Vertex AI; monthly evaluation on newly closed complaints.
Limitations & bias notes
Mixed-language messages with heavy slang or abbreviations ("dah 3 hari x dpt refund") score lower confidence and are sent to manual triage below 0.70. Sarcasm is a known error source. Chinese and Tamil text is out of scope.
Ownership and sensitivity
Who approves access
- Owner, Contact Centre, Siti Hajar Ismail
Business-sensitive. Models and data products scoped to named business units.
Entitlement per business unit, approved by the owner; conditions attach.
- Customer PII
- Contains or processes personal data about customers (PDPA 2010).
Try it
Live sandboxFeedback
Details
- Updated
- 2026-07-18
- Latest version
- 2.2.0
- Licence
- Group Reuse
- Access
- Open to all BUs
- Framework
- Transformers
- Language
- Bilingual
Trained on
complaints-corpus →